Deep Field / case

LLM Hallucination

An LLM emits fluent, coherent text—citations invented, numbers wrong, facts uncoupled from external records—at throughput measured in tokens per second across serving nodes. Internally the generative process does not distinguish false from true tokens; both are confabulations priced against training loss on corpus patterns, not against domain Canon. Witness publishes voluminously; Canon that would compress publication into truth-bearing invariants was never installed at the architectural stratum required—only corpus-fidelity gradient descent enforces fit-to-substrate. What is maintained is coherent token-stream replication. Payment is inference compute under a corpus-fidelity objective, not enforcement, malpractice, or audit labour. Characteristic failure signature: fluent false output indistinguishable internally from fluent true output—Witness–Canon decoupling structurally identical to confabulation without a proxy-pricing check. Hallucination is not a bug; it is the diagnostic price of publication without renormalisation.